Research Of Image Quality Assessment Based On Structure Features

The distortion of image can not be avoided when image was collected,saved and transported.Therefore,it is very important to have a stable and reliable image quality assessment model to effectively monitor and feedback the quality of the image when it is destroyed.In addition,the image quality assessment model can distinguish which algorithm is better in image processing(such as image restoration).The image features which are conform to the Human Visual System(HVS)peculiarity were extracted in this paper,in order to build a stable and efficient frame.According to the theory of structural Similarity and the peculiarity that the phase information and gradient information of the image contains more structural features,we propose an new image quality assessment model named Image Quality Assessment based on Frequency Domain Feature and Gradient Feature,hereafter this paper will be abbreviated as FGSIM.In this method,the frequency domain feature is a linear combination of amplitude feature and phase feature.A large weight is used to highlight the importance of phase feature.According to the experiment,in the frequency domain,we find that: the higher the amplitude value is,the higher the contribution of the corresponding phase characteristics to the image quality.Therefore,the phase feature is weighted by the frequency domain amplitude spectrum data.A novel image quality assessment method is proposed.First of all,R,G,B three color components were calculated for the quality score,and then the three scores were weighted according to the color sensitivity function of the human eye to get the color mean opinion score.In addition,the saliency map is used to mix local score together.Due to the strong nonlinear fitting ability and self-learning ability of machine learning algorithm,we introduce generalized regression neural network to construct the map between image feature and mean opinion score.Firstly,the phase feature,amplitude feature and gradient feature are extracted from R,G and B.Secondly,comparing the three characteristics of the training image and the reference image,the similarity score of them is obtained.A generalized regression neural network is used to learn a mapping model of similarity value and objective evaluation score,and then the mean opinion score is calculated by the constructed model.This method is also known as Image Quality Assessment based on Generalized Regression Neural Network(GIQA).On the basis of aforementioned two method,in order to evaluate the effectiveness of proposed method,the experiment is conducted on the LIVE,CSIQ and TID2013,which are compared by many articles.Five performance indexes are presented in the part of experiment.Experimental results on test images demonstrate that the proposed method achieved very competitive performance compared with VIF,SSIM and FSIM.